Two-step and direct approaches in controls, autonomous kites and the gap between control theory and applications

When dealing with problems of prediction, estimation, or control of complex systems, two starting ingredients are usually available: prior information on the problem at hand and experimental data. Depending on how these two ingredients are employed, two families of solution approaches can be distinguished: two-step and direct ones. The former techniques involve a first step where a model of the system is derived from prior information and data, and a second step where the desired outcome, for example a filter or a control algorithm, is designed on the basis of the derived model. Direct approaches, on the other hand, aim to design the desired outcome in one step, directly from data, without deriving explicitly a model of the system. In this talk, the advantages and limits of two-step and direct techniques will be compared in light of their application to the field of airborne wind energy, a new technology for wind energy conversion that has been enabled by advances in sensing, real-time computing and control design. The considered two-step approaches will be a predictive controller and a hierarchical control system; as direct approach, a technique to learn a nonlinear controller from data will be introduced. Finally, the presented results will be used to support some considerations on the interplay between theoretical developments and practical applications of controls.